"The Instrument (Antique Brass) is used exclusively for interactive elements." That sentence is prose, not code, and it sits in a file called ALPINE_OBSERVATORY.md — one of the test fixtures inside the DESIGN.md repository at packages/cli/src/linter/fixtures/.
Three lines above it, in the same file's YAML frontmatter, sits the token: primary: #f6bb81.
The hex value tells an agent what color to use. The sentence tells it where it may use it. Neither is sufficient alone, and that pairing is the format's entire thesis.
DESIGN.md is a plain-text specification for describing a visual identity to coding agents. A file has exactly two layers: YAML frontmatter holding machine-readable design tokens, and a Markdown body holding human-readable rationale. The spec states the relationship flatly: "The tokens are the normative values; the prose provides context for how to apply them."
The tokens are the normative values; the prose provides context for how to apply them.
The Alpine Observatory fixture is built on 19th-century polar expedition aesthetics. It carries 48 color tokens, all named in Material Design roles — surface-container-lowest, on-primary-fixed-variant — plus typography and spacing groups. Its telemetry-label token is IBM Plex Mono at 10px with 0.2em letter spacing, a precise way of saying "this should look etched into a brass instrument."
That precision is the point. "Slightly wider letter spacing" is an instruction a model will interpret differently every time. 0.2em is not.
The Heritage fixture demonstrates why the token layer has to win when the two layers disagree. Its prose describes Primary as #1A1C1E Deep Ink. Its frontmatter says primary: #000101 and primary-container: #1a1c1e. The prose is a narrative; the tokens are the contract.
Vercel learned the same lesson the expensive way. In an August 2026 writeup, John Phamous described porting an internal design skill into a single public prompt. It failed. The prompt "described our visual language just fine," but every model reading it interpreted that description differently, generating vastly different pages from the same guidance.
The diagnosis was specific: inside their codebases, agents read design guidance surrounded by real components and shipped examples. A public prompt includes none of that, leaving every model to rebuild the style from words alone. Phrases like "keep the layout clean" can mean anything. What is clean?
So they rebuilt it from scratch and tested every change against seven fixed eval prompts — usage report, renewal proposal,